| Summary: | Controlling the quality of products is imperative for any company. Every time a product is not working as it should, it makes the costumer unsatisfied and, therefore, loose the company a lot of money. So, if one could make the quality control process much more efficient, it would be beneficial for the business and the company itself. In the case of Bosch, we will be working with data from the quality testing of Printed Circuit Boards from heating boilers produced in the company. The testing process consists in making measurements of any kind related to the board and checking if they belong to a pre-set interval. The board will pass the test if that happens and, if not, it fails. The goal of this project is to find labels to each type of error that appears in the dataset and use them to make recommendations of what could be the most adequate repair to the failure in question. So, throughout this report, initially we present a previous study of the behavior of the quality control tests done throughout time to, consequently, find patterns that help us classify the different types of failure and create a recommendation system that indicates the repairs for each type of failure. This will also be helpful to succeeding processes, such as the checking of the repair and the repair itself, since the technicians will have some background knowledge on what they can do. In practical terms, we will firrstly analyze the dataset in order to identify distinct behaviors and important information, such as the test steps that never fail and the ones that fail the most, to then use it to our favor. We will apply machine learning techniques - Random Forests and Gradient Boosted Trees - in order to predict the values and failures of a certain step. With this, we will label the different types of failures, via the Root Cause Analysis methodology and the creation of metrics related to each failure, based on their behavior through time, which will be key information to the building of a recommendation system - with Collaborative Filtering techniques - that recommends what repairs should be done on the different failures. This report was inspired in problems presented in the application for the project "Augmented Humanity" (Projeto Mobilizador nº46103), P2020, which was approved.
|
|---|